# Weight Diff

A weight diff is the numeric half of [Model Diff](/docs/core-concepts/model-diff): a per-element `|A − B|` comparison between two aligned tensors, rendered wherever comparison mode is available - the [Diff view](/docs/editors/comparison-editor), the Matrix/Heatmap Diff shading mode, and the Signal/Distribution Delta sub-mode.

## Always logical, never raw bytes

A weight diff always compares fully reassembled, logical tensor values - never raw per-shard or per-file bytes. A checkpoint re-saved with a different sharding strategy, a different file layout, or a different serialization format still diffs as unchanged if its logical values didn't change.

This is why a [sharding strategy change](/docs/core-concepts/shards-and-storage#why-storage-splitting-doesnt-affect-diffing) alone never shows up as a false difference, and why comparing a safetensors checkpoint against an ONNX export of the "same" model (once properly [aligned](/docs/core-concepts/model-diff#alignment-comes-first)) is a meaningful comparison rather than something that trivially fails on format grounds.

## Reading a diff heatmap

The diff shading mode uses Tensormorph's standard diverging color convention - cyan where the candidate increased relative to baseline, red where it decreased, dark where nothing changed - so a diff heatmap is readable using exactly the same color vocabulary as the Signed shading mode elsewhere, not a diff-specific palette you have to relearn.

## Diff statistics, not just the picture

Alongside the visual heatmap, a diff carries its own [tile statistics](/docs/inspect-and-analyze/tensor-statistics) - mean absolute difference, max difference, and the fraction of elements that changed at all - computed the same mergeable way as any other statistic, so "how different are these two tensors, roughly" is answerable at a glance from the Inspector panel without reading the whole heatmap.

## From numeric to semantic

A weight diff on its own only tells you *where* and *how much* changed. See [Model Diff](/docs/core-concepts/model-diff#semantic-diff) for the best-effort classification layer on top - distinguishing a difference that looks like quantization noise from one that looks like a genuine, meaningful change - and [Quantization Diff](/docs/compare-and-morph/quantization-diff) specifically for the quantized-vs-full-precision case.

## Related resources

-   [Model Diff](/docs/core-concepts/model-diff) - alignment and semantic classification underlying every weight diff.
-   [Structural Diff](/docs/compare-and-morph/structural-diff) - comparing architecture, for when shapes or module structure themselves differ.
-   [Comparison Editor](/docs/editors/comparison-editor) - every view a weight diff can be rendered in.